AI sovereignty starts on the computer you own

"AI sovereignty" usually comes up in conversations about nations and data centres. But the same question applies at a much smaller scale: to a founder, a researcher, a household or a ten-person company. Who controls the context your AI works from, and what happens to it when you change your mind?

For most people, the honest answer today is: not you. Your history lives inside whichever assistant you picked first. Your preferences are stored in a format you can't read. Switching providers means starting over.

We think sovereignty at this scale comes down to three practical questions.

1. Where does your context live?

The most valuable thing you give an AI isn't any single prompt. It's the accumulated picture of your work: the people, the decisions, the constraints, the things you've corrected it on.

In Manwe, that record lives on your Mac: conversations, Worlds, memory, advisors and saved work. API keys and supported connection secrets sit in the macOS Keychain. Learning signals, and the small classifiers trained on them, stay on your Mac too.

That doesn't mean nothing ever leaves. It means the source of truth is a machine you own, and what leaves is a deliberate, per-task choice.

2. Which models see what?

Here we want to be precise, because this is where vague privacy claims fall apart.

The model you choose determines where model processing happens. A local model runs on your Mac's Apple silicon. A cloud model receives the relevant task context, which can include selected memories, document content and tool results. If several advisors use different providers, the relevant context can go to more than one of them.

And choosing a local model doesn't make every tool offline. Research, connected services, maps, downloads and sharing still use the internet when you ask for them.

Sovereignty here doesn't mean pretending everything is local. It means you can see the boundary and choose where it sits, task by task: a local model for the sensitive draft, a frontier model for the hard reasoning, and your memory controls deciding what either of them sees.

3. Can you leave?

The quiet test of any tool is what happens when you stop trusting one part of it. If a provider changes its terms, its pricing or its quality, can you move without losing years of context?

This is the reason Manwe is a harness rather than a model. It runs the models you choose: on your Mac, through subscriptions you already have, or with any API key. Your Worlds, memory and advisors stay put when you switch. The model is a replaceable part. Your world isn't.

What this looks like in practice

• Keep a confidential World on a local model, and let a cloud model handle public research.
• Require review before a World keeps new memories, so nothing enters context without your approval.
• Inspect What Manwe sees for a World before a sensitive conversation.
• Turn personal context off for a single chat when it shouldn't apply.

None of this needs a policy team. It needs software that treats your context as yours by default, and is honest about where the edges are.

Sovereignty isn't about where the servers are. It's about whether you can still answer "who has my context, and could I take it back?" a year from now.